FPT SoftwareFHNCertificationsReference architectures
Talk to NGT
MLOps · 03 of 5

Release governance

Fail-closed gate, human approval and a signed manifest that traces every verdict.

From candidate to production
  1. Candidatetrained and tested, never deployed
  2. Evaluation gatefail-closed
  3. Shadowbeside the current model
  4. Sign-offa person approves
  5. Productionone-click rollback
01

From candidate to production

Every model takes the same path, and each step ends at a gate.

  1. 01Candidate
    • Trained and tested by an Airflow run
    • That run never deploys
    a candidate in the MLflow registryfixed benchmark
  2. 02Evaluation gate
    • PR-AUC, regression and per-slice checks
    • Fail-closed: no pass, no release
    three-tier approval
  3. 03Shadow
    • Runs beside the current model
    • Its verdicts compared, not used
    a person signs
  4. 04Production
    • Served by Triton
    • One-click rollback to the last signed model
02

A signed manifest for every release run

Each release run writes a digitally signed manifest that locks its inputs.

Locked inputsWhat it answers
Signed run manifestone per release run

Point at a source to see what it feeds, or at a result to see where it comes from.

No auto-promote, ever

A candidate that beats the benchmark still waits for a person. Automation proposes; people release.

Questions this raises

The questions this part raises, and their answers.

A better average can hide a worse result on one segment or class. The gate checks each slice, and a person who knows the business signs.

SourcesAI Modeling & MLOps case study, NGT (Sep 2026)AI-First DevOps Transformation: Services & Solutions, NGT (Oct 2026)NGT Service Offering v3.2 (Jul 2026)Machine Learning Platform with Kubeflow, NGT reference deckPlatform Engineering v1.2, NGT (Sep 2026)Data Platform Ops & DataOps Service Catalog, NGT (Sep 2026)FPT AI Factory service offering, FPT Smart Cloud (Nov 2025)